- Private Credit Draws Both Caution and Confidence for Insurers
Global lenders and rating agencies are taking sharply different views of insurers’ fast-growing exposure to private credit.
- Model Updates May Also Push Texas Reinsurance Costs Down Further After Legislative Intervention
TWIA's reinsurance costs are projected to drop further, driven by new legislation and declining modeled loss scenarios for hurricanes.
- Climate Policies Are Missing the Mark on Heat and Cold Mortality Risks
Most government climate policies focus on extreme heat, but new research shows that moderate temperatures drive the majority of temperature-related deaths.
- While Hannover Re Feels Client “Love,” Investors Probe Growth and Returns
Analysts pressed Hanover Re’s management on how its “pure-play” reinsurance model, and all that client “love”, will translate into measurable growth.
Artificial intelligence may be transforming weather forecasting, but according to Dr. Hansi Singh, founder of Planette.AI, it won’t replace physics anytime soon.
“We need to continue making our physics-based models better in order to have better data to train any kind of AI emulator,” Singh told the Risky Science Podcast. “The physics and the AI need to go hand in hand.”
Singh, an Earth system scientist who’s worked with the U.S. Department of Energy and now provides long-range forecasts to finance and insurance clients, says the current breakthroughs in AI forecasting have been limited to short-term weather prediction.
“AI has revolutionized the seven-day forecast,” she said. “But if you want to predict next month, or next season, you just don’t have enough training data from the real world to do that effectively.”
AI’s Limits and Long-Range Gaps
The seven-day forecast is a sweet spot because meteorologists have 50 years of hourly atmospheric reanalysis data to train on. Beyond that, the Earth’s memory fades: the atmosphere “forgets” its starting point after two weeks, making it impossible to model purely from observation.
Long-term prediction depends on ocean, ice, and land-surface interactions, Singh noted, and that requires physics-based models to generate the training data AI still lacks.
“It’s much easier if you nudge AI toward the right answer,” Singh explained. “And that answer is fundamentally physics.”
Resolution and Cost
Where AI does shine, Singh said, is in resolution.
Traditional physics-based models double in computational cost—and energy use—with each halving of grid size. “The cost increases by a factor of eight,” Singh said. “AI downscaling can take a 25-kilometer model and sharpen it to near-street level.”
That’s a potential breakthrough for insurers and reinsurers seeking property-level insight. “In order to get that kind of hyperlocal information, AI could be super helpful,” Singh said. “But the AI still needs to be trained on the physics.”
From Wildfires to Counterfactuals
Singh highlighted how AI is uncovering new physical linkages, like teleconnections between Arctic sea ice and California wildfire risk, and how those insights can improve seasonal risk modeling.
She also sees AI’s role in scenario testing:
“If you know the AI model is as skilled as the physics-based one, you can just run more ensemble members—more counterfactuals—because AI models are computationally efficient.”
Trust, Transparency, and the Black Box
For risk modelers and regulators wary of “AI on top of a black box,” Singh said the key is validation and openness.
“We merge AI forecasts with physics-based models based on how each has performed at that location and lead time,” she said. “And we publish five years of back-tests right on our platform.”
That transparency, she argues, is how insurers and investors can build trust in next-generation models that combine the statistical power of AI with the physical rigor of climate science.
“It’s not about replacing physics,” Singh said. “It’s about using AI to strengthen it.”
🎧 Listen to the full episode: available now on the Risky Science Podcast.
